at 25 birthday, i spent sometime at san diego beach and wrote my core principles, drawing from my experiences and text/interviews from @paulg and @RayDalio . here is what i wrote:
1. embrace curiosity
continuously explore new ideas, fields, and technologies. embrace the mistake learner's high and strive to understand the power that comes from knowing how you and others are wired.
2. choose hard problems
aim high and tackle challenging problems that interest you. remember, difficult problems call for great efforts and often lead to the most rewarding outcomes.
3. failure is feedback
don't fear failure, see it as an opportunity for learning and growth. embrace adaptation through rapid trial and error and get a version 1.0 out fast, then continue to improve.
One month ago I gave Claude a $100,000 portfolio and let it trade the S&P 500 on its own. No human override. It is down 3.1%, within half a point of the index, every fill published before the open and never deleted. The honest scorecard, live: https://t.co/fzf2b0kVYd
My AI trader (nightclaude [dot] com) has refused to buy for a week.
~90% in Treasuries, near-zero equity, even with the S&P near record highs. Real money, leverage available, and it just won't. Tonight: "about as defensive as it gets."
Flat on the month, +0.49% ahead of the S&P.
/ So I catalogued all of it into "finesse" — a skill your AI coding agent reads before it builds any UI.
I ran the experiment: same prompt, same model (Opus 4.8), max effort. One read the skill, one didn't.
Look at the difference: https://t.co/9TlMkD1exq
Repo: https://t.co/oSnkiYoN3e
1/ Most AI-generated UIs look almost right and feel cheap.
It's not the model. It's the ~200 invisible details great designers obsess over. I spent years and $$$ learning those from the best in the world. I put all of it into a free skill.
As a product manager I got to build alongside ridiculously talented designers from Uber, Stellic, Cash App, Airbnb, and OpenAI.
npx skills add arjunlohan/finesse
Works with Claude Code, Cursor, Codex, and more.
2/ And it's never one big thing. It's the accumulation:
→ easing that decelerates in, accelerates out
→ exits a beat faster than entrances
→ concentric border radius
→ tabular numbers that don't jump
→ a focus ring you can actually see
Invisible alone. Obvious together.
Every market night, NightClaude [dot] com tears apart one S&P 500 company.
Six analysts, a bull, a bear, a judge, and a fact-checker that audits every number against SEC filings. No human edits a word.
11 companies so far. The next one publishes tonight. Free.
Every night at midnight, an AI publishes an institutional-grade deep dive on one S&P 500 company. Not a summary. A full report: business model, moat, valuation, bear case, verdict.
Free. Forever.
Here's how a machine writes equity research 🧵
Which company should it cover tonight? Drop a ticker below; most-liked reply wins, and we'll tag you when your report publishes at midnight.
(archive in the reply below 👇)
retards - most "AI trading" products are paywalls over backtests no one can replicate.
So I built one in the open:
- Claude tunes the strategy nightly using Karpathy's autoresearch loop
- $100k portfolio
- Full methodology and code will be public, live equity curve
Launches Tuesday. https://t.co/9KBvORDiFE
nothing ruins a Sunday morning quite like your AI co-worker clocking out for the day exactly one click before pushing your code.
oh claude claude claude 🥺